[llvm] [SLP] Fix canConvertToFMA operand selection and fmul costing (PR #216425)

[Dmitry Sidorov via llvm-commits]

================
@@ -2,25 +2,71 @@
 ; RUN: opt -passes=slp-vectorizer -S -mtriple=amdgcn-amd-amdhsa -mcpu=gfx90a -slp-threshold=14 < %s | FileCheck %s
 ; RUN: opt -passes=slp-vectorizer -S -mtriple=amdgcn-amd-amdhsa -mcpu=gfx942 -slp-threshold=14 < %s | FileCheck %s
 ; RUN: opt -passes=slp-vectorizer -S -mtriple=amdgcn-amd-amdhsa -mcpu=gfx950 -slp-threshold=14 < %s | FileCheck %s
+; RUN: opt -passes=slp-vectorizer -S -mtriple=amdgcn-amd-amdhsa -mcpu=gfx90a -slp-threshold=12 < %s | FileCheck %s --check-prefix=THR12
+; RUN: opt -passes=slp-vectorizer -S -mtriple=amdgcn-amd-amdhsa -mcpu=gfx942 -slp-threshold=12 < %s | FileCheck %s --check-prefix=THR12
+; RUN: opt -passes=slp-vectorizer -S -mtriple=amdgcn-amd-amdhsa -mcpu=gfx950 -slp-threshold=12 < %s | FileCheck %s --check-prefix=THR12
 
-; Elementwise d = c + a * b, where the fmul is operand 1 of the fadd. The
-; threshold puts the decision right at the cost boundary, so how the fmul and
-; fadd are priced against a fused fma is what decides it. These targets halve
-; the cost of a packed fmul, which is what tempts SLP into vectorizing and
-; breaking the scalar fma chain.
+; Elementwise d = c + a * b, where the fmul is operand 1 of the fadd. These
----------------
MrSidims wrote:

The test to be removed/shorten after rebase once pre-commit patch is merged.

https://github.com/llvm/llvm-project/pull/216425